clawlike-code

clawlike-code

A persistent-memory plugin for Claude Code: per-session injection, full transcripts, end-of-session summaries, and a per-turn classifier that catches what the agent forgot to write down. Inspired by openclaw; aligned with the CoALA framework.

What it does

Four equal-weight features. All zero-config, all running on every session.

1. Memory injection (SessionStart)

Walks .clawlike/context/*.md, parses each file's contract: frontmatter, emits a <persistent-memory> envelope to the agent. Multi-file structure means routing edits to the right file isn't a guess — each file's contract declares what belongs.

2. Session transcription (Stop, every turn)

Writes every conversation to .clawlike/sessions/<session-id>.md — full markdown with ## User / ## Assistant headers, committable, greppable. Replaces ad-hoc transcript exports.

3. Session summarisation (Stop, final)

On the agent's last stop of a session, prepends a one-paragraph Haiku summary to the top of the session file. Future sessions can scan the headers to understand what happened previously without rereading the whole transcript.

4. Per-turn safety-net classifier (Stop, first)

After each turn, a fast Haiku call looks at the recent conversation + every memory file's contract and asks: "Did the agent state a durable rule and fail to write it down?" If yes, it:

• Writes the proposal to MEMORY.md ## Pending Lessons (timestamped, with target file + reasoning)
• Returns {"decision": "block", "reason": "..."} to Claude Code

Claude Code injects the reason as a synthetic meta-message and continues the agent loop — so the agent addresses the proposal in the same turn (apply it, reject it, or explain why not). The classifier is throttled by a 10-second cooldown.

Why

Most Claude Code memory plugins either dump everything into a single CLAUDE.md or stuff observations into an opaque vector DB. clawlike-code does neither:

• Multi-file, contract-driven. Each memory file's contract: frontmatter declares its purpose. The classifier reads contracts and proposes targeted edits — no overflow, no drift.
• CoALA-aligned. Semantic / procedural / episodic taxonomy, the convergent shape of the agent-memory field.
• Repo-native. Everything lives in markdown files in your repo. No daemons, no vector DBs, no external secrets. Git is the persistence layer.
• Zero-config. No API keys, no env vars, no setup file. Auth piggybacks on Claude Code's session ingress token, so cost is on whatever billing you're already using.

Install

/plugin install clawlike-code

Then in any Claude Code session:

/clawlike-code:init

That scaffolds .clawlike/context/ with the 6 canonical starter files. Fill in the body of each, commit, and you're done.

The 6 canonical files

Borrowed directly from openclaw's workspace convention:

File	Stores	CoALA type
IDENTITY.md	Name, role, what I am. Stable facts.	Semantic
SOUL.md	Voice, character, what I won't do. Not workflows.	Semantic
AGENTS.md	Hard rules, SOPs, always/never.	Procedural
USER.md	Who I work with — schedule, preferences, communication style.	Semantic
TOOLS.md	Toolbelt + local cheat-sheet.	Semantic
MEMORY.md	Incident-derived lessons + the classifier's pending-proposal inbox.	Curated episodic

Adding your own

The plugin walks .clawlike/context/*.md indiscriminately — any file with a contract: frontmatter is loaded. Useful additions worth considering:

• PHYSICS.md — the shape of the agent's runtime: sandbox, timeouts, persistence model. Helpful for agents that run on infrastructure with non-obvious constraints.
• MAP.md — external context sources, data locations, "where do I find X" lookup table. Per the AX framing: an agent's two jobs are use the map and improve the map.
• Topic-specific files — known-issues.md, relationships.md, experiments.md. The same contract: pattern; the classifier will route to them if their contracts cover the case.

The memory skill (auto-invoked on relevant prompts) teaches the agent how to choose between files and when to suggest adding new ones.

Architecture

Two hooks. Zero daemons. Zero vector DBs.

SessionStart hook
└─ lib/inject.sh
   ├─ walks .clawlike/context/*.md
   ├─ parses each file's contract: frontmatter
   └─ emits <persistent-memory> envelope (methodology + per-file <layer> blocks)